CAREER: Towards Hierarchical and Provably Safe Control for Learning-Enabled Autonomous Systems

这笔学院早期职业发展计划(CALEAR)补助金将用于资助研究，以提高依赖机器学习技术的自主系统的可靠性、可信度和社会接受度，从而促进科学进步，促进国家繁荣和福利。自动驾驶汽车和无人机等自动机器人系统正在塑造着这个国家的未来，就运输、物流和经济的服务部门而言。人工神经网络已经成为现代自主系统中不可或缺的组成部分，特别是在它们的感知和控制管道中。然而，神经网络是复杂的，难以分析，并且对输入扰动或对抗性攻击敏感。这使得他们严格的分析和设计非常具有挑战性。因此，尽管近年来继续保持乐观和巨大的技术进步，但由于突出的安全和可靠性问题，真正的自主系统仍然难以实现。该项目通过建立严格的方法框架和高效的算法来克服这些顾虑，用于分析、验证、运动规划和控制设计具有学习功能的组件的安全关键型动态系统。它展示了该框架如何为安全可靠的运行提供可证明的性能保证。通过研究、教育和推广的紧密结合，该项目旨在利用知识发现来刺激教学，利用灵感教学来鼓励对研究的兴奋，并使新产生的知识向公众开放。这是通过基于主动学习的机器人安全控制课程的设计，通过吸引来自代表性不足群体的学生参与研究并组织具有动手机器人活动的K-12暑期研讨会，以及通过提高公众对自主系统安全相关技术的识字、意识和信任来实现的。本研究旨在为具有神经网络组件的自主系统的多速率和被证明是安全的运动规划和控制建立严格的数学框架的基础。它通过研究基于约束区域和混合区域的算法来计算神经反馈系统的过近似可达集，并在计算效率和逼近精度之间进行可调折衷；基于稳健的二次规划方法来设计可证明安全的周期性事件触发跟踪控制器；基于二阶锥规划的具有连续时间安全保证的神经反馈系统的轨迹规划方法；以及在规划层和跟踪层之间具有假定保证合同的可证明安全的多速率规划和控制算法。理论结果的验证和确认将使用高保真车辆动力学软件模拟和在两个基于实验室的机器人平台上进行物理实验。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This Faculty Early Career Development Program (CAREER) grant will fund research that enhances the reliability, trustworthiness, and societal acceptance of autonomous systems that rely on machine learning-enabled technologies, thereby promoting the progress of science, and advancing the national prosperity and welfare. Autonomous robotic systems, such as self-driving cars and drones, are shaping the nation's future insofar as the transportation, logistics, and service segments of the economy are concerned. Artificial neural networks have become an indispensable component of modern autonomous systems, especially in their perception and control pipelines. However, neural networks are complex, difficult to analyze, and sensitive to input perturbations or adversarial attacks. This renders their rigorous analysis and design very challenging. Thus, despite the continued optimism and tremendous technological progress in recent years, truly autonomous systems remain elusive because of outstanding safety and reliability concerns. This project overcomes these concerns by establishing a rigorous methodological framework and efficient algorithms for the analysis, verification, motion planning, and control design of safety-critical dynamic systems with learning-enabled components. It demonstrates how this framework enables provable performance guarantees for safe and reliable operation. Through close integration of research, education, and outreach, the project aims to leverage knowledge discovery to stimulate teaching and learning, use inspired teaching to encourage excitement in research, and make newly generated knowledge accessible to the public. This is accomplished through active learning-based design of a course on safety control in robotics, by engaging students from underrepresented groups in research and organizing K-12 summer workshops with hands-on robotics activities, and by increasing public literacy, awareness, and trust in safety-related technologies for autonomous systems.This research aims to develop the foundations of a mathematically rigorous framework for the multi-rate and provably safe motion planning and control of autonomous systems with neural network components. It achieves this aim by investigating constrained zonotope- and hybrid zonotope-based algorithms for computing over-approximated reachable sets for neural feedback systems with a tunable trade-off between computational efficiency and approximation accuracy; robust quadratic program-based methods for designing provably safe, periodic event-triggered tracking controllers; second-order cone program-based trajectory planning methods for neural feedback systems with continuous-time safety guarantees; and provably safe multi-rate planning and control algorithms with an assume-guarantee contract between the planning and tracking layers. Verification and validation of the theoretical results will be performed using high-fidelity vehicle dynamics software simulations and with physical experiments on two lab-based robotic platforms.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.